Source-linked AI summary
Sensorless damage-safe grasping
Yusei Shuto, Danilo Vasconcellos Vargas
TL;DR
The paper asks whether encoder and motor-effort signals alone can support adaptive grasping of soft objects. Its strain-bounding controller achieves ≥98% grasp at 0% damage across ε=0.7–2.0% at k=4000 N/m, where neither fixed-force baseline does.
Problem
Whether encoder and motor-effort data already available on actuated grippers can support adaptive soft-object grasping is an open question, while dedicated tactile sensing is used for that purpose.
Source: §1 · p. 1Method
The controller estimates compression strain from effort-based force and a known stiffness lower bound, then closes to a user-set strain limit with no force sensor or per-fruit calibration.
Source: §1 · p. 1Results
≥98% grasp at 0% damage across ε=0.7–2.0% at k=4000 N/m; neither fixed-force baseline reaches this result.
Source: §5.2 · p. 6Takeaways & Limitations
The strain bound gives operators a conservative damage-control parameter, while closing speed trades throughput against gentleness.
Source: §7 · p. 9In the very-soft regime, rigid-pad grasping requires compression near the damage line; compliant grippers remain the standard approach.
Source: §6.2 · p. 7
Abstract
from arXiv · showhide
Robotic fruit harvesting must hold produce securely without bruising it, yet compression stiffness varies several-fold with ripeness within a single species, so no fixed grip force spans the range. Rather than tune force, we bound deformation: a controller closes the gripper until the object's estimated compression strain reaches a user-specified limit $\varepsilon$, using only the encoder position and motor-effort signal on every servo gripper---no tactile or force-torque sensor. Dividing an effort-based contact force by a lower bound on object stiffness makes the stop provably conservative---true compression stays at or below $\varepsilon$---for any $\varepsilon$ above a contact-detection strain floor we identify and quantify: robust detection itself spends compression, linearly in closing speed, making speed an explicit throughput--gentleness knob. Unlike a hand-tuned force threshold, $\varepsilon$ is a certified, size-scaling, operator-interpretable damage limit, and a ready safe-action parameter for learned grasping policies. In MuJoCo simulation over a realistic fruit-stiffness range, under a sensor-noise model calibrated to the real servo, the controller holds $\ge 98\,\%$ grasp at $0\,\%$ damage across all medium-to-firm stiffnesses for the entire certified $\varepsilon$ range, which neither fixed-force baseline attains; on stiffness-graded 3D-printed TPU cubes it matches baseline grasp success at roughly half the grip force and cuts soft-object damage from $100\,\%$ to $40\,\%$.
1. Introduction
This work uses encoder and motor-effort signals to bound grasp deformation by a user-specified compression strain, with a conservative guarantee above a quantified contact-detection floor. Simulation and hardware validation across varying stiffnesses show lower damage than fixed-force grasping at comparable grasp success.
1. Introduction
- Dedicated tactile or vision–tactile sensing adds cost, fragility, and integration complexity, while prior approaches use heuristic damage handling or fixed force without a damage model.The work asks whether motor-effort and encoder data already available on actuated grippers can support adaptive soft-object grasping.Source: §1 · p. 1, Figure 1 · p. 1
- The controller stops closing when effort-based force divided by a lower bound on object stiffness reaches the user-specified strain bound ε, using only encoder and motor-effort signals.It requires no tactile or force-torque sensor and no per-object calibration.Source: §1 · p. 1
- True compression remains at or below ε for any ε above the contact-detection floor, given only a lower bound on object stiffness.The guarantee is conservative across the fruit-stiffness range and requires no per-object calibration.Source: §1 · pp. 1–2
- Noise-robust contact detection itself spends compression linearly with closing speed, setting a minimum attainable ε and making speed a throughput–gentleness knob.The quantified floor also explains the reported hardware observations.Source: §1 · p. 2
- Simulation and hardware validation across varying object stiffnesses show lower damage than fixed-force grasping at comparable grasp success.The simulation uses a sensor-noise model calibrated to the hardware.Source: §1 · p. 2
2. Related Work
Prior work addresses fruit damage through gripper design and tactile sensing, while sensorless methods estimate force or grasp state without providing a user-tunable compression-strain limit. The related work also identifies contact-detection compression as an unquantified cost of proprioceptive grasping.
2.1. Robotic fruit harvesting
- Harvesting systems use rigid, cutting, suction, soft, pneumatic, or tactile/slip-aware grippers, yet variable fruit properties leave preset grips unreliable and damage-free grasping open.Source: §2.1 · p. 2
2.2. Tactile and vision sensing for deformable grasping
- Han et al. fuse GelSight and RGB vision to label deformable grasps safe, slippery, or damaging, with success of 38 % on bananas and 90 % on round fruit; related fusion also needs costly tactile hardware [13].Source: §2.2 · p. 2
2.3. Sensorless force estimation from motor current
- Sensorless methods estimate force or grasp state from actuator signals, but do not provide a user-bounded compression-strain limit; this work uses a conservative stiffness lower bound and quantifies detection compression as speed-dependent.Prior force-estimation work uses generalized momentum and motor current, disturbance observers or learned friction models, deep temporal networks, gripper calibration, and finger proprioception; deformable-object methods predict poses or deformation properties, while stiffness is often probed or learned with uncertainty.Source: §2.3 · p. 2
3. Method
The method uses encoder position and motor effort to estimate compression strain and control jaw motion toward a user-set bound. Dividing the effort-based force proxy by a stiffness lower bound makes the estimated strain conservative, so the stop bounds true deformation.
3.1. Overview
- At 50 Hz, the controller uses jaw position and motor effort alone to estimate strain and command jaw velocity, without tactile or force-torque sensing.Before a grasp, the operator supplies ε, object diameter D, and stiffness lower bound k_min; closing speed v trades throughput against gentleness, and D is the main per-object assumption.Source: §3.1 · pp. 2–3, Figure 2 · p. 3
3.2. Contact detection
- Contact detection uses an amplitude threshold rather than a rate detector, with a 10-sample effort baseline and three-tick persistence.The threshold is set at 3–4σ of free-closing noise; until detection latches, the controller closes at full velocity and ignores the force proxy, while detection itself spends compression and sets a lower bound on attainable ε.Source: §3.2 · p. 3
3.3. Force proxy and stiffness lower bound
- The effort signal provides a force proxy that is divided by a fixed stiffness lower bound rather than per-fruit stiffness measurements.The motor-specific effort-per-force coefficient need not be calibrated precisely; an online Hookean stiffness fit was disabled because safe stops provided too little displacement for reliable identification.Source: §3.3 · pp. 3–4
3.4. Compression-strain estimate
- The online strain estimate is r_hat = F_hat/(k_min D), using the effort-based force proxy, stiffness lower bound, and object diameter.It estimates compression depth as a fraction of undeformed diameter because the controller cannot measure compression depth directly without tactile sensing.Source: §3.4 · p. 4
3.5. Conservative compression-strain bound
- The stiffness lower bound makes estimated strain over-bound true strain, so stopping at ε certifies true compression at or below ε.The guarantee is attainable only above the compression spent detecting contact; the stiffness margin also absorbs force-calibration error when the effort-coefficient error stays within the stiffness ratio.Source: §3.5 · p. 4
3.6. Stop policy
- A proportional close-or-hold law reduces jaw velocity as estimated strain approaches ε and re-closes during lifting if strain drops.With release velocity set to zero, overshoot clips the action to zero; no dead-band is used because it suppressed lift re-closing and reduced grasp success.Source: §3.6 · p. 4
3.7. Baseline controllers
- The fixed_low and fixed_high baselines differ only in their force thresholds and share the proposed method’s contact detector and lift trajectory.fixed_low stops on contact; fixed_high uses a threshold chosen for a firm grip on the stiffest specified object. Neither uses diameter or the stiffness lower bound, so their stop thresholds do not encode ε.Source: §3.7 · p. 5
4. Experiments
The experiments evaluate the controller in MuJoCo and on stiffness-graded TPU cube proxies, using calibrated motor-effort contact detection and explicit grasp, strain, damage, and timeout criteria.
4.1. Simulation
- Simulation sweeps six strain limits across five fruit-stiffness settings, with 50 seeds per cell and calibrated effort-signal noise at the hardware operating point.The MuJoCo setup uses 35, 40, or 45 mm cubes, stiffnesses of 2–10 × 10^3 N/m, and 5 mm/s closing speed.Source: §4.1 · p. 5
4.2. Hardware setup
- Hardware trials compare three strain limits with two fixed-force baselines on 40 mm TPU cubes at soft, medium, and stiff infill levels.The 75 trials use five seeds per method–cube cell; the authors note that validation on real fruit remains future work.Source: §4.2 · p. 5
4.3. Evaluation metrics
- Grasp success uses a slip-floor current proxy on hardware and an end-state slip check in simulation; damage is flagged above 1.5% strain or by visual score.In simulation, strain must exceed 1.5% for at least five consecutive ticks; on hardware, an operator score of at least 1 flags damage. A 500-tick failure to commit is a timeout.Source: §4.3 · pp. 5–6
5. Results
Across the stiffness sweep, the proposed controller achieves at least 98% grasp success with 0% damage in the medium-to-firm range, while fixed-force baselines fail to reach that result consistently. On hardware, it reduces soft-cube damage relative to both fixed-force baselines, while contact-detection latency sets a lower limit on achievable strain.
5.1. Simulation
- 0.35–0.56% settled strain across the grid for ε ≤ 1.3% shows that contact-detection latency, not ε, sets the stop in this regime.Because ε = 0.3% lies below the detection floor, the controller cannot stop below the strain already spent detecting contact, and the resulting grips are too weak to lift reliably.Source: §5.1 · p. 6
5.2. 3D-Printed Objects
- 40% damage on the soft TPU cube contrasts with 100% damage for both fixed-force baselines.At medium and stiff cubes, the proposed method matches the baselines at 100% grasp and 0% damage in every cell except 80% grasp at ε = 0.7% on the stiff cube, using roughly half the grip force. On the soft cube, the fixed-force baselines time out in 60% and 100% of trials, while the proposed method has 0% timeout.Source: §5.2 · p. 6
- At k = 2000 N/m, even stop-on-contact bruises 70% of cubes, so every fixed force threshold damages at least 70% of the softest objects.The proposed controller remains damage-free at ε ≤ 1.4%; raising a fixed threshold to recover grasp success at k = 4000 N/m still bruises at least 70% of the softest cubes.Source: §5.2 · p. 6
- ≥98% grasp at 0% damage is maintained for every stiffness k ≥ 4000 N/m and every ε ≥ 0.7%; neither fixed-force baseline reaches this across the sweep.The result holds in 17 of 30 cells at exactly 100% grasp and 0% damage; at k = 4000 N/m, the clean corner occurs at ε ∈ {0.7, 1.4, 2.0}%.Source: §5.2 · p. 6
- At k = 2000 N/m, damage stays at 0% for ε ≤ 1.4%, but grasp peaks at only 48%; at ε = 2.0%, damage reaches 76% while grasp is 18%.At ε = 1.5%, the trade-off is 40% grasp at 6% damage.Source: §5.2 · p. 6
6. Discussion
The discussion presents strain-bounded grasping as a damage-safe alternative to fixed-force control across medium-to-firm objects, while identifying a detection-cost floor and limits on very soft produce. It also frames the strain target as a safe parameter for learned grasping and outlines the hardware validation and stiffness-estimation work still needed.
6.1. What the method achieves
- At every k ≥ 4000 N/m and ε ∈ [0.7, 2.0]%, simulation achieved ≥98% grasp at 0% damage, while neither fixed-force baseline did.On hardware, the method matched baseline grasp success on medium and stiff cubes at roughly half the grip force, and reduced soft-cube damage from 100% to 40% at ε = 0.7%. Across the sweep, true strain exceeded ε in 0.8% of trials at ε ≥ 0.7%, all on the softest cubes.Source: §6.1 · p. 6
6.2. What the method does not do
- At k = 2000 N/m in simulation and on soft TPU cubes, no ε both secured a grasp and avoided damage; this is a rigid-pad limitation, not a controller failure.Soft-pneumatic or compliant grippers remain the standard approach for the very-soft regime.Source: §6.2 · pp. 6–7
6.3. The detection floor, and closing speed as its knob
- For ε ≤ 1.3%, stop strain was 0.35–0.56% across the grid because contact detection itself consumes compression before the controller engages.The floor rises linearly with closing speed; ε below that floor, including the 0.3% sweep row, is unattainable. Reducing speed during the persistence window can lower the floor without slowing the entire close, though hardware validation of this refinement remains future work.Source: §6.3 · p. 7, Figure 7 · p. 8
6.4. Sim-to-real consistency
- On soft hardware cubes, damage rose from 40% to 80% to 100% at ε = 0.7%, 1.3%, and 2.0%, respectively, mirroring simulation’s k = 2000 column.Without a deformation target, fixed-force baselines timed out and bruised every soft cube, while the strain controller stopped at ε and committed on every trial.Source: §6.4 · p. 8
6.5. Relation to fixed-force control
- The controller uses a certified, size-scaling force threshold derived from ε, k_min, and D, not online stiffness adaptation or a hand-picked force value.The bound holds when k_min ≤ k_true and ε clears the detection floor; during lifting, a proportional law re-closes when arm-joint yield drops. This enables the ≥98% grasp / 0% damage regime at k ≥ 4000 N/m that neither fixed-force preset attains.Source: §6.5 · pp. 8–9
6.6. A safe action space for learned grasping
- A learned grasping policy can choose ε above the detection floor as a pre-certified safe action when k_min lower-bounds the object class.The paper casts this as a one-dimensional safe exploration space, analogous to a shield in safe reinforcement learning [32], and notes that controller-parameter action spaces ease learning in contact-rich manipulation [33].Source: §6.6 · p. 9
6.7. Sim-to-real considerations and future work
- The hardware tests used stiffness-tunable TPU cubes, not biological fruit, whose anisotropy, viscoelasticity, and irregular geometry challenge the stiffness fit and single-diameter assumption.Online effective contact-dimension estimation is a proposed extension. Damage-safe stopping also suppresses the compression needed for reliable live stiffness estimation; brief probing or a learned contact-onset estimator could provide excitation, with stiffness potentially serving as a ripeness or quality cue.Source: §6.7 · p. 9
7. Conclusion
The paper presents a sensorless controller that bounds grasp deformation using a user-specified compression-strain limit, with a conservative stop above a measurable contact-detection floor. In simulation and TPU-cube tests, it combines damage-limited grasping with lower force or damage than fixed-force baselines.
7. Conclusion
- Simulation achieved ≥98% grasp at 0% damage from 4000 N/m upward across ε=0.7–2.0%, a regime neither fixed-force baseline attained; TPU cubes matched baseline success at roughly half force and reduced soft-cube damage from 100% to 40%.In the cube tests, the controller stopped at the target strain while fixed-force baselines kept closing and over-compressing.Source: §7 · p. 9
- The controller estimates compression by dividing an effort-based force proxy by a known lower bound on object stiffness, then closes until strain reaches user-specified ε.Because the stiffness is a lower bound, true compression stays at or below ε for any limit above the contact-detection strain floor; detection compression grows linearly with closing speed.Source: §7 · p. 9
- The stiffness margin absorbs motor-effort calibration error and supports hardware transfer without per-object force calibration, but rigid pads cannot reliably lift very soft objects while preserving a conservative damage bound.Future work includes validating on real fruit, relaxing the fixed-diameter assumption, and estimating stiffness online to tighten bounds on firm fruit and provide a ripeness cue.Source: §7 · p. 9